Documente Academic
Documente Profesional
Documente Cultură
269–301
DOI: 10.1111/j.1468-2443.2010.01113.x
ABSTRACT
This paper takes a fresh look at the importance of liquidity risk using a
comprehensive liquidity measure, weighted spread, in a Value-at-Risk (VaR)
framework. The weighted spread measure extracts liquidity costs by order size
from the limit order book. Using a unique, representative data set of 160
German stocks over 5.5 years, we show that liquidity risk is an important risk
component. Actually, liquidity risk is increasing the total price risk by over
25%, even at 10-day horizons and for liquid blue chip stocks and especially in
larger, yet realistic order sizes beyond h1 million. When correcting for liquidity
risk, it is commonly assumed that liquidity risk can be simply added to price
risk. Our empirical results show that this is not correct, as the correlation
between liquidity and price is non-perfect and total risk is thus overestimated.
I. INTRODUCTION
n
We would like to thank Deutsche Börse AG for providing the ‘Xetra Liquidity Measure (XLM),’
weighted spread for the major German stock indices. We are grateful to Tobias Berg, the
participants of the Munich Finance Seminars and the German Finance Association Annual
Meeting 2009 for helpful comments. The usual caveats apply. Comments or questions are highly
welcome. An earlier version of this paper was available under the title ‘Why and How to Integrate
Liquidity Risk into a VaR-Framework’ on October 30, 2008.
r 2011 The Authors. Journal compilation r International Review of Finance Ltd. 2011. Published by Blackwell
Publishing Ltd., 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148, USA.
International Review of Finance
analyze a data set of the Xetra Liquidity Measure (XLM) for 160 stocks over 5.5
years. Usual samples are restricted to few stocks over few months, because
weighted spread has to be calculated manually from the whole intraday order
book, which is highly computational extensive. Our sample is available from
Deutsche Börse.
With the help of this sample, we empirically analyze the magnitude of
liquidity impact at standard, larger than intraday horizons. We also clarify
whether tail correlation between price and liquidity costs is perfect or not.
The remainder of the paper is organized as follows. Section II defines
liquidity, the liquidity measure and liquidity risk and discusses the situations
when our approach is valid. In Section III, we describe our empirical data set
and the empirical results. Section IV summarizes and concludes.
In Section II.A, we first define liquidity from a cost perspective, characterize the
situational assumptions in which our framework can be applied and describe
our empirical liquidity measure. In Section II.B, we introduce our risk
estimation approach as well as a risk decomposition to distill structural insights.
where T(q) are direct trading costs, PIt(q) is the price impact versus mid-price
due to the size of the position and Dt(q) are delay costs if a position cannot be
traded immediately.
Direct trading costs T(q) include exchange fees, brokerage commissions and
transaction taxes.5 They will be neglected for further analysis as they are very
small and not relevant in a risk setting where time variation is of interest.6
Delay costs Dt(q) comprise costs for searching a counter party and the cost
imposed on the investor due to the bearing price risk and price-impact risk
during the execution delay.7 We assume that both forced and deliberate delays
are not relevant.
As we look at asset positions, which are continuously tradable during crises,
forced delay is not relevant. This means that no (or very few) zero trading days
occur in crises and the position size is not larger than the market depth.
Scanning our data, zero trading days appear to occur mainly in calmer market
periods, which, we hypothesize, happens because tumbling market prices
attract traders, which in turn ensures continuous trading.
We also assume that a deliberate, strategic delay has no significant benefit,
i.e. we assume that positions can be equally good when instantly liquidated
against the limit order book. This also neglects liquidation via limit instead of
market orders as well as up-floor or over-the-counter trading. So, we neglect any
(potential) effect of optimal trading strategies, which balance the increased
price risk of delay against the reduced liquidity cost by trading smaller
quantities (cp., e.g., Almgren and Chriss 1999, 2000; Almgren 2003; Bertsimas
and Lo 1998; and others). In our view, this is a reasonable assumption in four
cases. When we adopt the worst-case perspective of impatient traders, a
common risk assumption, potential benefits are consciously neglected. Benefits
are also non-existent if informational content of our trade is too high. The
trader wants to trade immediately on an informational advantage, which would
be revealed by trading more slowly or dissolve over time. Adverse informational
effects are also possible, i.e. trading more slowly could have price effects because
the market assumes informational advantage, which is not present in reality.8
Immediate liquidation is fair, too, if liquidity prices are efficient and a trader’s
risk aversion is greater or equal to that of the market.9 In this case, the marginal
gain from lower liquidity costs by delaying a transaction balances the marginal
loss due to higher price risk. Finally, optimal trading strategies might not be
feasible in times of market stress,10 because the optimization parameters are not
stable or strategic trading is not always possible.
As direct trading costs and delay costs are argued to be insignificant in our
analysis, we focus on price impact PIt(q) only, the difference between the
transaction price and the mid-price resulting from imperfectly elastic demand
and supply curves. For small volumes this is the bid–ask spread, but for larger
volumes the price impact is larger. Thus, if markets are fairly liquid, positions
are not too large and we adopt a worst-case perspective, the total liquidity cost
can be fairly measured with the price impact from immediate execution
(L(q)5PI(q)).
Price
= absolute liquidity cost = XLM(q) × q
a
P
b
q Order size
Quote depth / Size of next-best in €
Size of best limit orders bid limit order
12 Liquidity cost estimation could gain further precision if exchanges would provide buy-side- and
sell-side-weighted half-spread data.
13 Up to now, it been empirically impossible to distill precise price impact measures for single
assets from ex-post transaction data (cp. Amihud 2002; Pastor and Stambaugh 2003).
order book. The bid–ask spread is the minimum weighted spread (for small
order sizes). However, spread is usually not measured for constant order sizes q,
because quoted depth differs between stocks. Spread also has upper bounds
regulated by the exchange protocol if there is market maker coverage.14 This is a
first indicator that liquidity-cost dynamics will be different when moving
beyond the minimum bid–ask spread.
ask- and bid price at time t, respectively. rt;Dt is the Dt-period continuous mid-
price return at time t, i.e. rt;Dt ¼ lnðPmid;t =Pmid;tDt Þ . We adopt a traditional
approach from a VaR perspective and define price risk as the relative VaR at the
(1a)% confidence level over the horizon Dt
VaRa;Dt a
price ¼ 1 expðrt;Dt Þ ð5Þ
14 On Xetra illiquid stocks, defined by XLM and past volume, are covered by market makers. If the
stock is not covered, the bid–ask spread corresponds to the minimum spread in the order book.
15 Cp. Stange and Kaserer (2008) for a detailed discussion of the properties of XLM.
a
where rt;Dt is the a percentile of Dt-period return distribution. Consequently,
VaRprice measures the maximum percentage loss over the period Dt with a
confidence of (1a)%.
Analogously, we measure the total risk including the liquidity risk. To
calculate the impact of liquidity, we define the Dt-period net return in t as the
sum of the continuous mid-price return and the liquidity discount converted to
a continuous value
rnett;Dt ðqÞ ¼ rt;Dt þ lt ðqÞ ¼ ln Pmid;t =Pmid;tDt þ lnð1 Lt ðqÞÞ: ð6Þ
Note the difference of (6) to net-price returns, i.e. ln Pmid;t
ð1 Lt ðqÞÞ= Pmid;tDt ð1 LtDt ðqÞÞ Þ. Using net returns instead of net-price
returns, we implicitly assume that the liquidity cost of entering a position has
already been properly accounted for. While the quantified difference might not
be large, we believe that using net returns is conceptually closer to reality.
Price is then calculated as
Pnet;t ðqÞ ¼ Pmid;tDt expðrt;Dt þ lt ðqÞÞ ð7Þ
where Pnet;t ðqÞ is the achievable transaction price.
The Dt-period liquidity-adjusted total risk is then defined in a VaR framework as
the empirical a percentile of the net-return distribution.
VaRa;Dt a
total ðqÞ ¼ 1 expðrnett;Dt ðqÞÞ ð8Þ
VaRtotal is the maximum percentage loss due to mid-price risk and liquidation
cost over the period Dt with a confidence of (1a)%. This specification covers
the real dynamics of the net return on a certain stock position. It is practical but
also more general than existing approaches in the following ways:
1 We use a more precise liquidity measure than most papers by covering more
aspects of liquidity. Specifically, we account for the impact of order size on
liquidity. This extends the approach of Bangia et al. (1998, 1999), where
liquidity costs of any order size is proxied for with the bid–ask spread. As the
spread is valid only for very small order sizes, it is insufficient, which will be
demonstrated in the empirical section. The XLM measure is also more
precise than the ones used in Berkowitz (2000), Francois-Heude and Van
Wynendaele (2001) or Angelidis and Benos (2006), because it directly
measures liquidity costs in the limit order book instead of proxying for it.
2 As we take empirical percentiles instead of a parametric method, we avoid
any distributional assumption, especially on liquidity cost, such as in Giot
and Grammig (2005). Our approach will capture the non-normality of the
distribution as well, which is made possible by our large sample size. Figure 2
shows that the empirical net-return distributions can be far from normal.
3 Our approach takes percentiles of the net-return distribution and does not
treat price risk and liquidity separately. We look at the dynamics of net
returns that combines the mid-price-return dynamics and liquidity cost
dynamics. Instead of adding distribution percentiles of liquidity and price
risk separately, we acknowledge that liquidity cost and mid-price might not
Figure 2 Exemplary net-return histogram and fitted normal distribution for the
Comdirect stock in 2007.
This figure shows the exemplary histogram of net-return of the Comdirect stock for
trade-size of h100 thsd. and normal distribution based on its mean and variance in 2007.
with lat;Dt being the empirical percentile of the continuous liquidity discount.
This is the maximum percentage loss due to liquidity cost at a (1a)%
confidence level.
We can now apply a further decomposition of total risk and define the
correlation factor k(q) as a residual of
VaRtotal ðqÞ ¼ VaRprice þ VaRliquidity ðqÞ þ kðqÞ VaRliquidity ðqÞ: ð11Þ
Naturally, this is just a further decomposition of the liquidity impact
VaRliquidity
lðqÞ ¼ ð1 þ kðqÞÞ: ð12Þ
VaRprice
k(q) measures the tail correlation factor between mid-price return and liquidity
cost, the proportion of liquidity risk, that is diversified away due to tail
correlation. In this definition, the correlation factor is always non-positive,
k(q)0. If tail correlation is perfect, k(q) is zero and worst mid-prices and worst
liquidity costs can be added to get the total risk as is, for example, done in
Bangia et al. (1999). If there is some diversification between cost and price, k(q)
will become negative.
The liquidity impact l(q) contains the following conceptual components. First,
it contains the mean liquidity discount for the position of size q – in contrast to
other approaches. This is suitable as positions are usually valued at mid-prices
already neglecting mean liquidity costs. Second, it includes negative deviations
from the mean cost as measured by volatility and higher moments. Third,
possible diversification effects between price and liquidity are included and
liquidity risk is reduced. If liquidity cost and mid-prices have a less than perfect,
negative tail correlation (k(q)o0), a liquidity risk estimate based on the a
percentile of the liquidity cost distribution as in (10) will be incorrectly higher
than that when based on the net-return distribution as in (9).
continuous liquidation over a certain time period is valid. This is also the
situation where optimal trading strategies can be applied to maximize the net
sales proceeds. In our framework, we are looking at a situation characterized in
II.A.i, which justifies instant liquidation. If we look at impatient traders or
equivalently at the worst case, we do not allow for mitigation of some of the
liquidity cost by allowing continuous liquidation. In such a case, ‘orderly
liquidation’ needs to be more generally defined and our approach is suitable.
In the empirical part, Section III.A describes our data set and Section III.B
provides some market background to our analysis. Section III.C presents our
empirical results and Section III.D includes our robustness tests.
A. Description of data
Our sample consists of 5.5 years of daily XLM data (July 2002–January 2008) for
all 160 stocks in the four major German stock indices (DAX, MDAX, SDAX andf
TecDAX).16 Therefore, in total, we cover a market capitalization of approxi-
mately h1.2 trillion (as of January 2008), which represents the largest part of the
market capitalization in Germany. As far as we know, this is the most
representative sample on weighted spread available to academia.
We received XLM data for all days, where a stock was included in one of the
four indices.17 Daily values are calculated by Xetra as the equal-weighted
average of all available by-minute data points.18 XLM(q) comprises for each day
the weighted spread for 10 standardized order sizes q. Standardized order size
reach from h25,000 to h5 million in the DAX and from h10,000 to h1 million in
all other indices. In addition to XLM data, we obtained the day-closing bid–ask
spread s at the Xetra trading system from Datastream.
Three stocks were excluded from the analysis due to missing XLM or
Datastream data.19 We also had to eliminate 408 XLM observations, where
liquidity data were available outside the standardized volume class structure
described above, to ensure that our estimates remain representative in each
16 The DAX contains the 30 largest publicly listed companies in Germany (by free-float market
volume), the MDAX includes the subsequent 70 largest before March 24, 2003 and 50 largest
thereafter and the SDAX includes the following 50 largest. The TecDAX, introduced during the
sample period on March 24, 2003, comprises the 30 largest technology stocks.
17 Therefore, our sample is non-constant containing 275 different stocks, but only 160 stocks at
one point in time.
18 This comprises a maximum of 1060 measurements during continuous trading.
19 Procon Multimedia (in SDAX between October 2002 and March 2003) and Medisana (in SDAX
between December 2002 and March 2003). Data could not be obtained for Sparks Networks (in
SDAX between June 2004 and December 2005), because it was not available in Datastream
anymore.
volume class.20 These exclusions left 99.9% or 323,670 of the stock days in the
sample.
In total, our remaining sample contains 1.8 million observations for the 1424
trading days. We divide our total sample into four sub-samples, each containing
the stocks of one index.
B. Market background
As background to our analysis, Table 1 summarizes market conditions during
the sample period. Markets were bullish in the largest part of the sample period.
We also captured the downturns in the second half of 2002 and the first month
of 2008. Because of beginning and end-of-period declines, the overall return
was rather average at 8% p.a. Naturally, market capitalization increased similar
to returns. Market capitalization is several times larger in the DAX than in all
other indices. MDAX contained the second largest average market capitaliza-
tion stocks, followed by TecDAX and SDAX. Volatility exhibited a similar, but
reversed pattern than returns. Because of the bullish period, our sample is
probably rather positively biased.
Daily transaction volume increased considerably during the sample period,
which is already a plausible indicator for improving liquidity. Transaction
volume was largest in the DAX; in the other indices it was several magnitudes
smaller. Contrary to the general positive trend, transaction volume in the
TecDAX remained steady after its initiation in 2003 and exhibits a level slightly
lower than the MDAX. SDAX transaction volume was again several times
smaller than that in MDAX or TecDAX. The high diversity in transaction
volumes underlines the representativeness of our sample.
20 Less than 0.01% of all observations were available for connected periods of o7 days. We assume
that the automatic calculation routine of the Xetra computer was extended to non-standard
order sizes during trial periods.
281
Table 2 Liquidity impact on risk (VaR, 10-day, 99%)
282
l(q), VaR(10- Order size (in thsd. euro) Size
day, 99%) in % impact
of price risk Minimum 10 25 50 75 100 150 250 500 750 1000 2000 3000 4000 5000 All
DAX
Mean (%) 1 NA 1 1 NA 1 NA 2 4 NA 9 16 21 24 26 10 0.78nnn
Median (%) 1 NA 1 1 NA 1 NA 1 3 NA 6 11 19 20 25 3 0.78nnn
Standard 1 NA 0 0 NA 1 NA 2 4 NA 8 14 16 16 18 14 0.84nnn
deviation (%)
Observations 42,129 NA 42,710 42,710 NA 42,710 NA 42,710 42,706 NA 42,663 41,71639,97038,22536,343 412,463
MDAX
Mean (%) 2 2 2 3 4 5 7 11 15 19 21 NA NA NA NA 8 0.58nnn
Median (%) 2 1 2 2 3 4 5 9 11 14 17 NA NA NA NA 4 0.62nnn
Standard 2 3 3 5 4 5 6 9 18 35 63 NA NA NA NA 22 0.62nnn
deviation (%)
Observations 69,578 74,779 74,574 73,930 73,291 72,671 71,357 68,520 60,784 53,461 46,741 NA NA NA NA 670,108
SDAX
Mean (%) 9 6 7 10 13 16 20 23 22 22 30 NA NA NA NA 14 0.35nnn
Median (%) 3 3 4 7 8 9 11 14 19 20 23 NA NA NA NA 8 0.44nnn
Standard 52 8 11 16 20 29 44 48 17 15 34 NA NA NA NA 27 0.23n
deviation (%)
Observation 69,988 69,081 64,254 60,824 57,798 54,871 49,291 39,780 23,114 13,985 8,363 NA NA NA NA 441,361
TecDAX
Mean (%) 3 1 2 3 4 5 8 11 16 18 18 NA NA NA NA 7 0.64nnn
Median (%) 1 1 1 2 3 4 6 8 13 18 16 NA NA NA NA 3 0.66nnn
Standard 5 1 1 2 4 8 7 10 18 12 13 NA NA NA NA 10 0.66nnn
International Review of Finance
deviation (%)
Observations 35,741 37,133 37,133 37,126 37,075 36,949 36,299 33,958 26,995 20,641 16,031 NA NA NA NA 319,340
All
Mean (%) 4 NA 3 5 NA 7 NA 12 13 NA 17 NA NA NA NA 10 0.44nnn
Median (%) 2 NA 2 2 NA 3 NA 7 10 NA 13 NA NA NA NA 4 0.62nnn
Standard 30 NA 7 10 NA 16 NA 25 16 NA 42 NA NA NA NA 20 0.42nn
deviation (%)
Observations 217,436 NA 218,671214,590 NA 207,201 NA 184,968153,599 NA 113,798 NA NA NA NA 1,843,272
Table shows cross-sectional statistics of lambda, which is the impact of liquidity in percent of price risk according to equation (9); minimum column
measures risk at minimum spread level; all column is average over all standardized order sizes, i.e. without minimum; size impact is the coefficient of
log-size regressed on the log distribution statistic including an intercept.
n
10%, nn5% and nnn1% confidence level of being different from zero based on a two-tailed test.
On average, over all stocks and across all order sizes, total risk – including
liquidity risk – is 10% higher than price risk alone. DAX is generally the index
with the lowest liquidity risk, while MDAX and TecDAX are second. SDAX
consistently shows the highest liquidity impact levels across all order sizes. This
finding is consistent with trading volumes and market values discussed in
Section III.B.
There is significant variation in the liquidity impact between indices and
within indices as indicated by standard deviations. Variation is of the same
order of magnitude than the level. Impact is practically zero (1%) in small
order sizes of the DAX (oh250 thsd.). Liquidity impact can easily increases
above 20% in large stock positions of the DAX or medium stock positions in
small stocks. In an average h1 million SDAX positions, liquidity impact on risk
increases to 30% of price risk at a 10-day horizon.
Especially interesting is the liquidity impact calculated with spread as
revealed in the minimum column. This corresponds to the risk measurement
approach suggested by Bangia et al. (1999). Impact remains rather small across
all stocks and comparable to the liquidity impact measured with XLM(10) and
XLM(25), respectively. In SDAX and TecDAX, it is slightly higher than in the
smallest XLM bracket. Because median risk levels are comparable, this effect is
probably due to few outliers as XLM and spread data come from two different
databases.
Liquidity impact generally increases with order size.21 To analyze this size
effect more systematically, we separately estimated the impact of doubling
order size on l(q) in percent in the last column. To do so, we regress the log row
statistics on log order size including a constant intercept.22 Size impact is the
coefficient on log size and indicates the curvature of the price-impact function.
It specifically investigates into the importance of price-impact data in contrast
to spread data only and abstracts from the different levels in liquidity risk
between indices. Generally, the estimated price-impact statistic is positive but
smaller than one, which shows that the liquidity impact (risk) function is
concave.23 The price impact is larger in the DAX than in the other indices. Here,
the difference between small, liquid and larger less-liquid positions is especially
pronounced. With a size impact of 0.78, liquidity impact almost doubles in the
DAX when doubling order size. In the other indices, liquidity impact is already
large at small positions – hence the lower curvature. All size impacts are
statistically significant at the 1% level. The economically large size-impact
21 The decrease in the average SDAX position between h250 thsd. and h500 thsd. results from a
non-constant sample effect. Large SDAX positions were continuously tradable only in later
years. Therefore, risk estimates for large SDAX positions are calculated on a more liquid period
depressing liquidity impacts compared with more continuously traded small positions.
22 Ordinary least-squared regression equation is logðStatðqÞÞ ¼ c þ logðqÞ þ e, with stat being the
row statistic and c being a constant intercept.
23 This is consistent as already the price impact cost function is empirically found to be concave
(cp. Hasbrouck 1991; Hausman et al. 1992).
statistic underlines the importance of using order book information beyond the
spread for risk estimation – even in the DAX.
These results have important consequences for risk estimation techniques.
First, we find that liquidity is an important component in total risk, especially
in larger order sizes, where the price-impact estimation error relative to price
risk rises up to 30% at 10-day horizons. Second, estimating liquidity risk with
spread data is no valid alternative, as liquidity risk impact in this size class is
very small and increases considerably with size. Third, large variations indicate
that the constant scaling of price risk across all stocks, ‘hair cuts,’ are probably
insufficient and liquidity has to be accounted for specifically for each stock.
DAX
Mean (%) 64 NA 40 41 NA 43 NA 45 49 NA 34 41 40 40 41 41 0.00
Median (%) 68 NA 41 41 NA 41 NA 44 48 NA 49 47 42 41 41 44 0.00
Standard 21 NA 15 15 NA 16 NA 17 18 NA 87 47 30 24 19 36 0.00
deviation (%)
Observations 42,129 NA 42,710 42,710 NA 42,710 NA 42,710 42,706 NA 42,663 41,716 39,970 38,225 36,343 412,463
MDAX
Mean (%) 70 62 62 61 63 64 63 63 64 60 59 NA NA NA NA 62 0.00nn
Median (%) 72 66 66 62 63 67 64 61 62 63 60 NA NA NA NA 63 0.00nn
Standard 15 17 17 18 16 16 17 16 18 24 17 NA NA NA NA 18 0.00
deviation (%)
Observations 69,578 74,779 74,574 73,930 73,291 72,671 71,357 68,520 60,784 53,461 46,741 NA NA NA NA 670,108
SDAX
Mean (%) 67 61 65 65 66 66 64 60 57 59 54 NA NA NA NA 63 0.00nnn
Median (%) 68 61 65 65 65 66 64 59 56 61 57 NA NA NA NA 64 0.00nn
Standard 19 19 16 18 16 16 18 15 17 18 16 NA NA NA NA 17 0.00
deviation (%)
Observations 69,988 69,081 64,254 60,824 57,798 54,871 49,291 39,780 23,114 13,985 8,363 NA NA NA NA 441,361
TecDAX
Mean (%) 72 66 66 64 66 66 67 66 63 65 62 NA NA NA NA 65 0.00nn
deviation (%)
Observations 35,741 37,133 37,133 37,126 37,075 36,949 36,299 33,958 26,995 20,641 16,031 NA NA NA NA 319,340
All
Mean (%) 68 NA 59 59 NA 61 NA 59 59 NA 50 NA NA NA NA 58 0.00nn
Median (%) 70 NA 61 61 NA 61 NA 58 58 NA 56 NA NA NA NA 60 0.00nn
Standard 18 NA 19 19 NA 18 NA 18 19 NA 56 NA NA NA NA 25 0.00nn
deviation (%)
Observations 217,436 NA 218,671 214,590 NA 207,201 NA 184,968 153,599 NA 113,798 NA NA NA NA 1,843,272
Table shows cross-sectional statistics of the correlation factor, which measures correlation between liquidity cost and mid-price return according to (11); minimum column
measures the effect at minimum spread level; all column is average over all standardized order sizes, i.e. without minimum; size impact is the coefficient in 102 of size in a
linear regression of the distribution statistic on size including an intercept.
n
10%, nn5% and nnn1% confidence level of being different from zero based on a two-tailed test.
285
International Review of Finance
However, the overall effect of this assumption remains small if the liquidity
impact is small in total. It might also be different in other assets like currencies,
which were analyzed by Bangia et al. (1999), but we see no a priori reason why
this should be the case. We also hypothesize that correlation effects should be
similar for other liquidity cost measures, because they proxy for the same
phenomenon. Overall, our results indicate that tail correlation is important and
should be taken into account in illiquid stock positions.
D. Robustness tests
i. Effect of using the expected shortfall (ES) measure
Recently, literature has discussed coherent risk measures as an alternative to VaR
to overcome the shortfalls of VaR like non-sub-additivity (cp. Artzner et al.
1997; Acerbi and Scandolo 2007). This raises the question as to whether our
results would change significantly when switching to a different risk measure.
To test whether our results are robust or specific to the VaR, we calculate ES, also
called ‘conditional VaR’ or ‘expected tail loss,’ which is the expected loss in the
DAX
Mean (%) 3 NA 2 2 NA 3 NA 5 11 NA 30 56 68 76 79 32 0.82nnn
deviation (%)
Observation 35,741 37,133 37,133 37,126 37,075 36,949 36,299 33,958 26,995 20,641 16,031 NA NA NA NA 319,340
All
Mean (%) 15 NA 13 18 NA 29 NA 43 52 NA 59 NA NA NA NA 36 0.43nnn
Median (%) 7 NA 7 9 NA 13 NA 33 43 NA 53 NA NA NA NA 17 0.61nnn
Standard 60 NA 23 31 NA 54 NA 65 50 NA 53 NA NA NA NA 52 0.22n
deviation (%)
Observations 217,436 NA 217,548 213,653 NA 206,264 NA 183,967 152,727 NA 113,237 NA NA NA NA 1,833,877
Table shows cross-sectional statistics of lambda, which is the liquidity impact on risk in percent of price risk according to equation (9); minimum column
measures risk at minimum spread level; all column is average over all standardized order sizes, i.e. without minimum; size impact is the increase in risk in
percentage points when doubling order size, measured as coefficient in 102 of log-size in a regression of the log distribution statistic on log-size including an
intercept.
n
10%, nn5% and nnn1% confidence level of being different from zero based on a two-tailed test.
287
International Review of Finance
worst a% of the cases. We continue to use our basic approach detailed in Section
II.B.ii, but we replace VaR with ES defined as follows:
ESa;Dt ¼ Eðrjr < r a Þ: ð13Þ
Similarly, we calculated liquidity-adjusted total risk with the mean mrnet and
variance srnet of a 20-day backward net-return distribution
VaR1%
total ðqÞ ¼ 1 expðmrnet ðqÞ 2:33srnet ðqÞÞ ð15Þ
with net returns calculated according to equation (6). We then calculate the
liquidity impact l(q) according to equation (9). Neglect of negative skewness
and high kurtosis (fat tails) makes this procedure simple, but might under-
estimate risk. Because of the underestimation, absolute values need to be treated
with care, but are still – as lower bound – a suitable indicator for the time
variation of the liquidity impact on risk, especially if higher moments are fairly
constant.
Results for l(q) on the basis of a 10-day, 99% VaR according to (9) and (15) are
illustrated in Figure 3, details shown in Table 6. The impact of liquidity on risk
has generally declined over time across all indices. In all years, the liquidity
impact increased considerably with order size as the size-impact statistic reveals.
Our previous finding of the index rank (DAX, MDAX/TecDAX and SDAX) is
confirmed and stable over time. TecDAX, however, was shortly more liquid after
its initiation in 2003 until 2004. Although to be interpreted with care, the
liquidity impact probably remained non-negligible during the low-risk period
DAX
Standard 2 1 1 2 3 4 6 8 11 16 27 NA NA NA NA 9 0.77nnn
deviation (%)
Observations 169 347 333 329 334 313 300 261 196 153 107 NA NA NA NA 2673
All
Mean (%) 3 NA 3 3 NA 5 NA 8 11 NA 15 NA NA NA NA 8 0.48nnn
Median (%) 1 NA 1 2 NA 3 NA 5 7 NA 10 NA NA NA NA 3 0.52nnn
Standard 7 NA 7 5 NA 13 NA 9 12 NA 25 NA NA NA NA 16 0.34nn
deviation (%)
Observations 1980 NA 2075 1994 NA 1887 NA 1575 1312 NA 926 NA NA NA NA 16,339
Table shows cross-sectional statistics of lambda, which is liquidity impact in percent of price risk according to equations (9) and (13); minimum column
measures risk at minimum spread level; all column is average over all standardized order sizes, i.e. without minimum; size impact is the coefficient in 102 of
log-size in a regression of the log distribution statistic on log-size including an intercept.
n
10%, nn5% and nnn1% confidence level of being different from zero based on a two-tailed test.
289
International Review of Finance
50%
45%
40%
35%
30%
25%
20%
15%
10%
5%
DAX 100
0%
2002 2003 2004 2005 2006 2007 2008
from 2006 to 2007. The impact of liquidity on total risk was certainly
economically significant in the crises periods of 2002–2003 and in 2008.
Results for the whole panel (‘all’) have to be treated with care, because they
are distorted by the non-constant sample effect. Over the years, the liquidity of
less-liquid stocks improved considerably, which made their liquidity cost data
increasingly available. As consequence, less-liquid, high-cost stocks are
increasingly included in the sample, which increases the average risk estimate.
However, individual year estimates have almost no sample bias and underline
the fact that liquidity impact is economically significant.
If skewness and kurtosis would be included, these findings are also likely to
get confirmed, as the one-time liquidity cost deduction will probably introduce
additional skewness, which keeps the relation between price and liquidity risk
valid. Overall, this confirms the fact that liquidity price impact is economically
significant enough to encourage integration into risk measurement systems.
2008 1 1 1 1 1 2 2 3 6 10 13 NA NA NA NA 4 0.67nnn
All 2 2 3 3 4 6 8 12 20 24 26 NA NA NA NA 10 0.61nnn
D2002–2008a 2 3 4 4 5 6 7 7 8 6 4 NA NA NA NA 2
SDAX
2002 13 12 27 61 91 120 153 115 79 NA NA NA NA NA NA 36 0.56nn
2003 10 12 17 26 36 42 40 41 40 51 64 NA NA NA NA 27 0.32nnn
2004 6 7 11 20 27 35 40 52 46 38 120 NA NA NA NA 24 0.49nnn
2005 6 5 7 11 15 19 24 29 32 41 37 NA NA NA NA 17 0.46nnn
2006 3 4 5 7 10 13 19 28 37 36 37 NA NA NA NA 15 0.56nnn
2007 2 2 3 4 5 7 10 17 31 38 43 NA NA NA NA 13 0.71nnn
2008 2 2 2 3 5 6 10 16 26 32 39 NA NA NA NA 10 0.75nnn
All 6 6 9 14 17 21 23 28 34 38 43 NA NA NA NA 18 0.42nnn
291
D20022008a 7 6 18 47 74 99 130 86 45 13 21 NA NA NA NA 18
292
Table 6 (continued)
Avg. l(q), Order size (in thsd. euro) Size
VaR(10-day, impact
99%) in % Minimum 10 25 50 75 100 150 250 500 750 1000 2000 3000 4000 5000 All
of price risk (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%)
TecDAX
2002 NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA
2003 2 3 4 5 8 11 15 21 25 28 31 NA NA NA NA 11 0.57nnn
2004 2 3 4 5 8 12 18 26 26 29 31 NA NA NA NA 13 0.60nnn
2005 2 2 3 5 6 8 11 17 27 38 41 NA NA NA NA 12 0.67nnn
2006 2 2 2 3 4 5 7 13 22 27 31 NA NA NA NA 10 0.68nnn
2007 1 1 2 2 3 4 5 8 15 19 23 NA NA NA NA 8 0.66nnn
2008 1 1 1 2 2 3 4 6 11 15 16 NA NA NA NA 5 0.65nnn
All 2 2 3 4 6 8 11 16 22 27 29 NA NA NA NA 11 0.62nnn
D2002–2008a 0 1 1 1 2 4 4 5 3 1 1 NA NA NA NA 1
All
2002 6 NA 8 10 NA 13 NA 12 14 NA 15 NA NA NA NA 13 0.14nnn
2003 5 NA 7 9 NA 14 NA 19 21 NA 21 NA NA NA NA 15 0.31nnn
2004 3 NA 5 8 NA 14 NA 20 22 NA 23 NA NA NA NA 15 0.42nnn
2005 3 NA 4 5 NA 9 NA 14 19 NA 24 NA NA NA NA 11 0.53nnn
2006 2 NA 2 3 NA 6 NA 12 16 NA 19 NA NA NA NA 9 0.59nnn
International Review of Finance
2007 1 NA 2 2 NA 4 NA 8 14 NA 19 NA NA NA NA 7 0.70nnn
2008 1 NA 1 2 NA 3 NA 7 11 NA 13 NA NA NA NA 6 0.69nnn
All 3 NA 4 6 NA 9 NA 14 17 NA 20 NA NA NA NA 11 0.45nnn
D2002–2008a 3 NA 4 4 NA 4 NA 2 3 NA 5 NA NA NA NA 2
Table shows mean lambda, which is liquidity impact in percent of price risk by sub-sample calculated with a rolling mean-variance estimation of
Value-at-Risk (10-day, 99%) according to (9) based on (15). Minimum column measures risk at minimum spread level; all column is average over
all standardized order sizes, i.e. without minimum; size impact is the coefficient in 102 of log-size in a regression of the log distribution statistic
on log-size including an intercept.
a
Statistic shows absolute change between 2003 and 2008 when 2002 number not available.
n
10%, nn5% and nnn1% confidence level of being different from zero based on a two-tailed test.
liquidity impact l(q) based on a 10-day, 99% VaR according to (9). While our
methodology does not use optimized position weights, a value-weighted
portfolio should show effects of diversification if there are any. Results are
presented in Table 7. Estimates demonstrate that the liquidity impact on the
portfolio level is of a similar magnitude as that on the average individual stock
level (cp. Table 2). Especially in larger sizes, liquidity impact is increased at the
portfolio level, for example, it increases to 54% for the h1 million position in
the SDAX portfolio compared with 30% for the average individual stock
position. This must be driven by larger liquidity commonality in larger sizes, i.e.
diversification in liquidity between stocks decreases with larger sizes. Even for
the all-stock portfolio, liquidity impact levels are higher than for the average
stock. Overall, our results are robust to diversification effects in stock portfolios.
In this paper, we modeled liquidity risk based on the weighted spread liquidity
measure in a VaR framework. The main advantage over existing approaches is
the higher precision of the weighted spread, which calculates liquidity cost
differentiated by order size, i.e. the price impact, from the limit order book.
We argued that weighted spread is a valid liquidity measure from a risk
perspective in a wide range of situations, which we defined clearly. If we look at
limit order book markets, where this type of data is available and from the
perspective of institutional investors, for whom other direct trading costs are
negligible, two situational assumptions are critical.
First, our approach works most precise for continuously tradable asset
positions, for example, for small- to medium-sized positions in developed stock
markets. Positions cannot be too large, like block holdings, and markets have to
be fairly liquid with few zero trading days. If this is not the case, forced
execution delay can incur costs that we have neglected.
Second, we assume that deliberate, strategic delay has no significant benefit,
which renders optimal trading strategies useless. This is a fair assumption in
four possible cases. We can adopt a worst-case perspective, for example, because
external restrictions require to close whole positions immediately. Any possible
benefit from delay is then consciously ignored. From a risk perspective, strategic
delay also remains with unrealizable benefit if the optimal trading strategy is
non-stable in crises situations and can therefore not provide any benefit on an
expected basis. Further, if liquidity prices are efficient in fairly liquid markets,
strategic delay has per definition no marginal benefit. Finally, optimal trading
strategies are also useless if the real or perceived (i.e. adverse) informational
content of the trade is high and delay only increases the probability of adverse
price movements.
This discussion shows that these cases cover a variety of situations. Overall,
our approach is most valid for up to medium-sized positions in generally
continuously trading markets.
DAX
Estimate (%) 1 NA 1 1 NA 1 NA 1 2 NA 12 22 18 25 27 10 0.84nnn
Observations 42,129 NA 42,710 42,710 NA 42,710 NA 42,710 42,706 NA 42,663 41,759 40,002 38,294 36,388 412,652
MDAX
Estimate (%) 2 2 2 2 3 5 9 5 13 21 20 NA NA NA NA 7 0.60nnn
Observations 73,279 74,858 74,679 74,040 73,365 72,737 71,409 68,543 60,788 53,501 46,793 NA NA NA NA 670,713
SDAX
Estimate (%) 5 32 15 27 32 52 56 26 39 45 54 NA NA NA NA 35 0.16n
Observations 70,048 69,197 64,614 61,119 57,938 55,000 49,410 39,920 23,442 14,435 9,112 NA NA NA NA 444,187
TecDAX
Estimate (%) 1 1 2 2 3 4 5 8 13 15 17 NA NA NA NA 6 0.61nnn
Observations 36,980 37,157 37,157 37,150 37,099 36,973 36,323 34,028 27,077 20,868 16,291 NA NA NA NA 320,123
All stocks
Estimate (%) 3 NA 2 2 NA 3 NA 4 5 NA 20 NA NA NA NA 6 0.55nn
Observations 181,212 NA 219,160 215,019 NA 207,420 NA 185,201 154,013 NA 114,859 NA NA NA NA 1,847,675
International Review of Finance
Table shows portfolio statistics of lambda, which is the impact of liquidity in percent of price risk according to (9); minimum column measures risk at minimum spread
level; all column is average over all standardized order sizes, i.e. without minimum; size impact is the coefficient of log-size regressed on the log distribution statistic
including an intercept.
n
10%, nn5% and nnn1% confidence level of being different from zero based on a two-tailed test.
REFERENCES
Acerbi, C., and G. Scandolo (2007), Liquidity Risk Theory and Coherent Measures of
Risk. Working Paper. Available at SSRN: http://ssrn.com/abstract=1048322.
Almgren, R., and N. Chriss (1999), ‘Value under Liquidation’, Risk, 12, 61–3.
Almgren, R., and N. Chriss (2000), ‘Optimal Execution of Portfolio Transactions’,
Journal of Risk, 3, 5–39.
Almgren, R. F. (2003), ‘Optimal Execution with Nonlinear Impact Functions and
Trading-Enhanced Risk’, Applied Mathematical Finance, 10, 1–18.
Amihud, Y. (2002), ‘Illiquidity and Stock Returns: Cross-Section and Time-Series
Effects’, Journal of Financial Markets, 5, 31–56.
Amihud, Y., and H. Mendelson (2006), ‘Stock and Bond Liquidity and its Effect on
Prices and Financial Policies’, Financial Markets and Portfolio Management, 20, 19–32.
Angelidis, T., and A. Benos (2006), ‘Liquidity Adjusted Value-at-Risk Based on the
Components of the Bid-Ask Spread’, Applied Financial Economics, 16, 835–51.
Artzner, P., F. Delbaen, J.-M. Eber, and D. Heath (1997), ‘Thinking Coherently’, Risk,
10, 68–71.
Bangia, A., F. X. Diebold, T. Schuermann, and J. D. Stroughair (1998), Modeling
Liquidity Risk With Implications for Traditional Market Risk Measurement and
Management. Working Paper No. 6, Financial Institutions Center at The Wharton
School.
Bangia, A., F. X. Diebold, T. Schuermann, and J. D. Stroughair (1999), ‘Liquidity on the
Ouside’, Risk, 12, 68–73.
Berkowitz, J. (2000), ‘Breaking the Silence’, Risk, 13, 105–8.
Bertsimas, D., and A. W. Lo (1998), ‘Optimal Control of Execution Costs’, Journal of
Financial Markets, 1, 1–50.
Buhl, C. (2004), Liquidität im Risikomanagement. PhD Thesis, University of St. Gallen.
Coppejans, M. T., I. H. Domowitz, and A. Madhavan (2001), Liquidity in an Automated
Auction. Working Paper. Available at SSRN: http://ssrn.com/abstract=295765.
Cosandey, D. (2001), ‘Adjusting Value at Risk for Market Liquidity’, Risk Magazine,
October 115–8.
Deutsche Boerse (2004), Xetra Release 7.1, Marktmodell Aktien. Technical Report
Deutsche Boerse (2005), Orderbuchstatistik (order book statistics xetra close-1.xls).
Available at http://www.deutsche-boerse.de
Deutsche Boerse (2008), Preisverzeichnis für die Nutzung der Handels-EDV der
Frankfurter Wertpapierbörse. Available at http://www.deutsche-boerse.de
Domowitz, I., O. Hansch, and X. Wang (2005), ‘Liquidity Commonality and Return
Co-Movement’, Journal of Financial Markets, 8, 351–76.
Dowd, K. (2001), Beyond Value at Risk – The New Science of Risk Management. Hoboken,
NJ: Wiley & Sons.
Francois-Heude, A., and P. Van Wynendaele (2001), Integrating Liquidity Risk in a
Parametric Intraday VaR Framework. Working Paper. 7th Belgian Financial Research
Forum, Liège.
Giot, P., and J. Grammig (2005), ‘How Large is Liquidity Risk in an Automated Auction
Market?’, Empirical Economics, 30, 867–87.
Gomber, P., and U. Schweickert (2002), ‘Der Market Impact: Liquiditätsma im
elektronischen Wertpapierhandel’, Die Bank, 7, 485–9.
Gomber, P., U. Schweickert, and E. Theissen (2004), Zooming in on Liquidity. Working
Paper. Available at SSRN: http://ssrn.com/abstract=559406.
Hasbrouck, J. (1991), ‘Measuring the Information Content of Stock Trades’, The Journal
of Finance, 46, 179–207.
Hausman, J. A., A. W. Lo, and A. C. MacKinlay (1992), ‘An Ordered Probit Analysis of
Transaction Stock Prices’, Journal of Financial Economics, 31, 319–79.
Irvine, P. J., G. J. Benston, and E. Kandel (2000), Liquidity Beyond the Inside Spread:
Measuring and Using Information in the Limit Order Book. Working paper.
Available at http://ssrn.com/paper=229959
Jarrow, R., and P. Protter (2005), Liquidity Risk and Risk Measure Computation.
Working paper, Cornell University. Available at http://people.orie.cornell.edu/
~protter/WebPapers/riskmeasliq6.pdf.
Jorion, P. (2001), Value at Risk: The New Benchmark for Managing Financial Risk, 2nd edn.
New York: McGraw-Hill, New York.
Jorion, P. (2007), Value at Risk: The Benchmark for Controlling Market Risk, 3rd edn.
McGraw-Hill Publishing Co.
Le Saout, E. (2002), Incorporating Liquidity Risk in VaR Models. Working Paper.
Available at: http://www.scribd.com/doc/7174003/Incorporating-Liquidity-Risk-in-
VaR-Models#stats.
Lei, C. C., and R. N. Lai (2007), The Role of Liquidity in Value at Risk – The Case of
Hong Kong. Working Paper. Available at SSRN: http://ssrn.com/abstract=1009578.
Loebnitz, K. (2006), Market Liquidity Risk: Elusive no more – Defining and quantifying
market liquidity risk. Diploma Thesis, University of Twente.
Pastor, L., and R. F. Stambaugh (2003), ‘Liquidity Risk and Expected Stock Returns’,
Journal of Political Economy, 111, 642–85.
Stange, S., and C. Kaserer (2008), The Impact of Order Size on Stock Liquidity – A
Representative Study. CEFS Working Paper 2008 No. 9. Available at http://ssrn.com/
abstract=1292304
APPENDIX A
This table contains distribution statistics on price risk calculated as 10-day, 99% VaR according to
equation (5).
This table contains distribution statistics on price risk calculated as 1-day, 99% VaR according to
equation (5).
DAX
Mean (%) 17 NA 17 17 NA 17 NA 17 17 NA 18 20 20 21 21 18 0.05nnn
Median (%) 15 NA 15 15 NA 15 NA 15 16 NA 17 18 18 17 18 17 0.04nnn
Standard deviation (%) 4 NA 5 5 NA 5 NA 5 5 NA 5 8 7 7 8 6 0.11nnn
Observations 42,129 NA 42,710 42,710 NA 42,710 NA 42,710 42,706 NA 42,663 41,71639,97038,22536,343 412,463
MDAX
Mean (%) 18 19 20 20 20 20 20 22 23 22 24 NA NA NA NA 21 0.05nnn
Median (%) 17 17 17 17 17 17 18 19 21 20 20 NA NA NA NA 18 0.05nnn
Standard deviation (%) 8 9 9 9 9 9 9 9 10 10 11 NA NA NA NA 9 0.03nn
Observations 73,234 74,779 74,574 73,930 73,291 72,671 71,357 68,520 60,784 53,461 46,741 NA NA NA NA 670,108
SDAX
Mean (%) 21 20 21 22 24 24 25 27 28 30 35 NA NA NA NA 24 0.11nnn
Median (%) 19 18 18 20 21 21 22 23 26 29 30 NA NA NA NA 21 0.12nnn
Standard deviation (%) 8 7 7 8 9 9 9 10 9 10 16 NA NA NA NA 9 0.12nnn
Observations 70,048 69,081 64,254 60,824 57,798 54,871 49,291 39,780 23,114 13,985 8363 NA NA NA NA 441,361
TecDAX
Mean (%) 25 21 22 22 22 22 23 24 26 26 30 NA NA NA NA 23 0.07nnn
Median (%) 23 19 20 20 20 20 21 22 24 24 28 NA NA NA NA 20 0.08nnn
Standard deviation (%) 9 8 8 9 8 9 9 9 10 9 10 NA NA NA NA 9 0.05nnn
International Review of Finance
Observations 36,980 37,133 37,133 37,126 37,075 36,949 36,299 33,958 26,995 20,641 16,031 NA NA NA NA 319,340
All
Mean (%) 20 NA 20 20 NA 21 NA 22 23 NA 23 NA NA NA NA 21 0.04nnn
Median (%) 18 NA 18 18 NA 19 NA 20 21 NA 21 NA NA NA NA 19 0.05nnn
Standard deviation (%) 8 NA 8 8 NA 9 NA 9 9 NA 11 NA NA NA NA 9 0.08nnn
Observations 222,391 NA 218,671214,590 NA 207,201 NA 184,968153,599 NA 113,798 NA NA NA NA 1,843,272
This tables shows cross-sectional statistics on empirical, absolute total risk including a liquidity adjustment according to equation (8); minimum column
measures risk at minimum spread level; all column is average over all standardized order sizes, i.e. without minimum; size impact is the increase in risk in
percentage points when doubling order size, measured as coefficient in 102 of log-size in a regression of the distribution statistic on log-size including an
intercept.
n
10%, nn5% and nnn1% confidence level of being different from zero based on a two-tailed test.
Observations 36,980 37,133 37,133 37,126 37,075 36,949 36,299 33,958 26,995 20,641 16,031 NA NA NA NA 319,340
All
Mean (%) 7 NA 7 7 NA 8 NA 9 10 NA 11 NA NA NA NA 9 0.12nnn
Median (%) 7 NA 7 7 NA 7 NA 8 9 NA 9 NA NA NA NA 7 0.10nnn
Standard deviation (%) 4 NA 3 3 NA 5 NA 6 5 NA 7 NA NA NA NA 5 0.24nnn
Observations 222,391 NA 218,671214,590 NA 207,201 NA 184,968153,599 NA 113,798 NA NA NA NA 1,843,272
This tables shows cross-sectional statistics on empirical, absolute total risk including a liquidity adjustment according to equation (8); minimum column
measures risk at minimum spread level; all column is average over all standardized order sizes, i.e. without minimum; size impact is the increase in risk in
percentage points when doubling order size, measured as coefficient in 102 of log-size in a regression of the distribution statistic on log-size including an
intercept.
n
10%, nn5% and nnn1% confidence level of being different from zero based on a two-tailed test.
299
Table A5 Liquidity-adjusted total risk on 10 day horizon (rolling estimate, VaR, 10-day, 99%)
300
L-adj. VaR Order size (in thsd. euro) Size
(10-day, impact
99%) in % Minimum 10 25 50 75 100 150 250 500 750 1000 2000 3000 4000 5000 All
(%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%)
DAX
2002 24 NA 24 24 NA 24 NA 24 25 NA 26 28 29 30 31 26 1.32nnn
2003 16 NA 17 17 NA 17 NA 17 17 NA 18 20 21 21 22 18 0.94nnn
2004 9 NA 9 9 NA 9 NA 10 10 NA 10 11 12 12 13 10 0.70nnn
2005 8 NA 8 8 NA 8 NA 8 8 NA 8 9 9 9 10 9 0.29nnn
2006 9 NA 9 9 NA 9 NA 10 10 NA 10 10 10 10 11 10 0.29nnn
2007 10 NA 10 10 NA 10 NA 10 10 NA 10 11 11 11 11 11 0.22nnn
2008 16 NA 15 15 NA 15 NA 15 15 NA 15 15 16 16 16 15 0.18nn
All 12 NA 12 12 NA 12 NA 12 12 NA 13 13 13 14 14 13 0.37nnn
D2002–2008a 12 NA 12 12 NA 12 NA 12 13 NA 14 15 16 16 17 14 0.96
MDAX
2002 21 21 21 22 22 23 24 26 28 27 27 NA NA NA NA 23 1.69nnn
2003 16 16 16 16 16 16 17 19 21 23 24 NA NA NA NA 17 1.87nnn
2004 11 12 12 12 12 12 13 13 16 17 19 NA NA NA NA 13 1.46nnn
2005 11 11 11 11 11 11 11 11 13 14 15 NA NA NA NA 12 0.80nnn
2006 12 12 12 12 12 12 13 13 13 14 15 NA NA NA NA 13 0.59nnn
2007 13 13 13 13 13 13 14 14 14 15 15 NA NA NA NA 14 0.49nnn
2008 20 19 19 20 20 20 20 20 20 21 22 NA NA NA NA 20 0.50nnn
International Review of Finance
All 14 14 14 14 14 14 14 15 16 16 17 NA NA NA NA 15 0.65nnn
D2002–2008a 7 7 7 8 8 8 9 11 13 11 10 NA NA NA NA 8 1.12
SDAX
2002 24 23 27 35 41 47 47 32 4 NA NA NA NA NA NA 28 1.45
2003 19 19 20 21 23 25 28 31 35 28 29 NA NA NA NA 22 2.98nnn
2004 14 15 15 16 18 20 21 25 31 31 33 NA NA NA NA 18 4.49nnn
2005 13 13 13 13 14 15 16 18 24 27 30 NA NA NA NA 15 3.83nnn
2006 14 14 14 15 15 15 16 18 23 27 31 NA NA NA NA 17 3.53nnn
2007 15 15 15 15 15 16 16 17 20 23 26 NA NA NA NA 17 2.22nnn
2008 20 22 22 22 21 21 22 24 28 33 35 NA NA NA NA 23 2.80nnn
All 16 16 16 16 17 17 18 20 23 25 28 NA NA NA NA 18 2.59nnn
2008 20 NA 20 20 NA 20 NA 21 22 NA 23 NA NA NA NA 21 0.84nnn
All 14 NA 14 15 NA 15 NA 16 16 NA 17 NA NA NA NA 15 0.72nnn
D2002–2008a 8 NA 8 9 NA 9 NA 9 10 NA 9 NA NA NA NA 9 0.30
Table shows liquidity-adjusted total risk by sub-sample according to equation (9) calculated with a rolling mean-variance estimation. Minimum
column measures risk at minimum spread level; all column is average over all standardized order sizes, i.e. without minimum; size impact is the
coefficient in 102 of log-size in a regression of the distribution statistic on log-size including an intercept.
a
Statistic shows absolute change between 2003 and 2008 when 2002 number not available.
n
10%, nn5% and nnn1% confidence level of being different from zero based on a two-tailed test.
301